# Prediction-market longshots are priced right, not too high

> In resolved binary prediction markets, longshots at 10c or less were off by only -1.1 to -1.6c at 1 to 14 days out; a +3 to +9c underpricing in a volume-selected sample vanished in an unbiased one.

E013 · Verdict: KILL · Cause: No signal · Market: Prediction markets · Mechanism: Behavioural · Horizon: 1d-14d · Tested: October 2026

Source: https://alphanotfound.eu/experiments/E013-prediction-market-longshot-bias

The favourite-longshot bias says bettors overpay for unlikely outcomes and underpay for likely ones. If so, selling a 5c longshot should pay on average. We checked 1,063 resolved on-chain prediction markets, then repeated the test on 1,072 markets sampled without regard to volume. The first sample showed longshots underpriced by 3 to 9c, the reverse of the bias. In the second, that was gone: longshots at 10c or less were priced about right (-1.1 to -1.6c). Selling longshots lost 2 to 8c per share in the holdout, after gaining at most 3c in the two earlier windows.

## Key numbers

- **Longshot mispricing:** -1.1 to -1.6c (realised minus price, p <= 0.10, every horizon (unbiased sample))
- **Volume-selected sample:** +3 to +9c (longshots looked underpriced; an artifact of how markets were chosen)
- **Sell-longshot holdout:** -2 to -8c (per share after costs, 1c and 2c half-spread)
- **Markets in unbiased sample:** 1,072 (454 events, 250 per month, volume floor 2k USD)

## The idea being tested

**Belief:** Unlikely outcomes are overpriced in betting markets, so systematically selling longshots (or buying favourites) earns money.

**Why people believe it:** People believe that bettors overpay for small chances (a 5c contract that wins less than 5% of the time) and underpay for near-certainties, and that the same behavioural bias would show up in prediction markets.

**Why it would persist:** The usual belief is that selling a longshot earns only a few cents per share against a rare large loss, so the bias would not be competed away.

## How it was tested

For each resolved yes/no market we read the Yes price 1, 3, 7 and 14 days before the market closed, while it was still open. We grouped prices into bins and compared each bin's price with how often Yes actually happened. Then we simulated two simple rules to resolution: buy No when Yes is at 10c or less, buy Yes when it is 90c or more. The first sample (the 3,000 most-traded markets) gave a surprising answer, so we repeated everything on markets chosen only by closing month.

**Pre-registered:** no

**Data:** Resolved binary markets with Yes price history, first sample: top 3,000 by traded volume, 1,063 markets in 745 events usable (an on-chain prediction market, public market and price-history API; price history at 12-hour fidelity; not stated for the volume-selected sample); Second sample: 1,072 closed binary markets in 454 events, 250 per month chosen by creation order among markets ending in that month, volume floor 2k USD (an on-chain prediction market, public market and price-history API; price history at 12-hour fidelity; end dates 2025-01..2026-09)

**Universe:** Binary markets only. Observation at h = 1, 3, 7 or 14 days before the end date, only if the market was still open then, the latest history point was at most 24 hours old and 0.02 < p < 0.98.

**Signal:** Yes price p at the observation time. Sell longshot: buy No when p <= 0.10 (bins from 0.02). Buy favourite: buy Yes when p >= 0.90 (up to 0.98). Both held to resolution.

**Execution:** One share per observation, held h days to resolution (longer than one day; reported for information).

**Costs:** Half-spread 1c primary, 2c stress; fee 0.07 x price x (1 - price) only in markets with fees enabled.

**Controls:** The calibration table is its own control: a calibrated price gives realised minus price of zero. The outcome-independent second sample is the control for the volume-selected first sample.

**Sample split:** By end date, 60% selection, 20% validation, 20% holdout. Standard errors are clustered by event.

**Pass criterion:** Reconstructed after the fact, not pre-specified: realised Yes frequency below price for longshots by more than the 1-2c half-spread and fee, with a sell-longshot rule positive after costs in selection, validation and holdout.

**Kill criterion:** Reconstructed after the fact, not pre-specified: no rule positive after costs in all three windows, or the mispricing smaller than the cost.

**Power:** Small: the favourite rule had 0 to 19 trades per window in the first sample, and 8 to 36 per window in the second. Longshot rule 9 to 190 trades per window and horizon.

## Result

**Why it was rejected:** In the outcome-independent sample, longshots are priced about right (-1.1 to -1.6c), about the size of the 1c half-spread, and no rule is positive in all three windows.

**Observations:** 75

**Net mean:** -3.64

**95% interval:** -11.74 to +4.47

**Median:** —

**Worst:** —

- In the volume-selected sample (1,063 markets, 745 events), realised Yes frequency exceeded price by +3 to +9c for p between 0.02 and 0.35, with event-clustered standard errors of 2.5 to 3.5c (t about 2 to 2.5), strongest at 7 days.
- Selling longshots in that sample mostly lost in selection and validation, -3.13c to -8.48c per share at 1, 3 and 7 days (1c half-spread); the one exception was +4.47c at 14 days in validation.
- The favourite rule in that sample was +4c in selection but had only 3 to 19 trades per window, so it says nothing.
- In the outcome-independent sample (1,072 markets, 454 events), longshots at p <= 0.10 were off by -1.1 to -1.6c at every horizon; bins 0.02 to 0.10 gave -1.7c (SE 0.8) and -0.7c (SE 1.5). Favourites were off by -0.1 to -3.4c with SE about 3c.
- In that sample, with a 1c half-spread, selling longshots netted -0.08c to +3.73c in selection and validation but -1.77c to -6.67c in the holdout (SE 2.6 to 7.1c); across the 1c and 2c half-spreads the holdout range is -2 to -8c. The holdout 7-day cell was -3.64c on 75 trades in 40 events; its 95% interval (-11.74 to 4.47) is the mean plus or minus 1.96 event-clustered SE, computed by us for one chosen cell, not a pre-specified test.
- The buy-favourite rule changed sign across windows; with a 1c half-spread it lost 13.20c, 9.38c, 6.77c and 5.86c per share in the holdout at 1, 3, 7 and 14 days.

### The sample choice flipped the sign (cents per share)

Selection-window net result of selling longshots (buy No at p <= 0.10, 1c half-spread). The volume-selected sample loses at every horizon; the outcome-independent sample is near zero or positive up to 7 days and about zero at 14.

|   | Top-volume markets (first sample) | Outcome-independent sample (second sample) |
| --- | --- | --- |
| 1 day | -4.2 | 1.48 |
| 3 days | -6.6 | 2.81 |
| 7 days | -8.48 | 2.77 |
| 14 days | -8.0 | -0.08 |

### Selling longshots fails in the holdout (cents per share)

Outcome-independent sample, 1c half-spread. Selection and validation are mostly small gains; the holdout is negative at every horizon, with standard errors of 3 to 7c.

|   | Selection | Validation | Holdout |
| --- | --- | --- | --- |
| 1 day | 1.48 | 3.73 | -6.67 |
| 3 days | 2.81 | 1.21 | -4.26 |
| 7 days | 2.77 | 1.09 | -3.64 |
| 14 days | -0.08 | 2.34 | -1.77 |

### Buying favourites changes sign by window (cents per share)

Outcome-independent sample, 1c half-spread. Positive and negative cells alternate in selection and validation, and the holdout is negative at every horizon (-5.86c to -13.20c).

|   | Selection | Validation | Holdout |
| --- | --- | --- | --- |
| 1 day | -1.12 | 3.83 | -13.2 |
| 3 days | 0.52 | -2.68 | -9.38 |
| 7 days | 1.34 | 4.28 | -6.77 |
| 14 days | -1.83 | -1.47 | -5.86 |

## Interpretation

**Why it failed (interpretation, not established):** The first sample picked the markets with the most trading volume. A likely explanation, not tested here, is that volume partly grows because the event happens: a longshot that starts to win attracts late money, so winning longshots would be over-represented and look underpriced without any mispricing. Once markets were chosen by closing month and a low volume floor, longshots were priced at or slightly above their realised frequency, and the gap was about the size of the cost of trading it.

**Lesson:** When testing resolved markets for calibration, any sampling variable that accumulates with the outcome (volume, for example) builds the answer into the sample. A bias that appears only in the most-traded subset is a property of the selection, so check it on a sample drawn before the outcome is known.

**Scope:** Covers resolved binary markets on one on-chain prediction market, prices 1 to 14 days before the end date, with end dates 2025-01 to 2026-09 in the second sample. It does not cover multi-outcome markets, prices closer than one day to resolution, or other venues.

## Reproducibility

**Steps:** 1) List closed binary markets from the public market API; for the second sample take 250 per month by creation order among markets ending that month with volume of at least 2k USD. 2) Download price history at 12-hour fidelity. 3) For h = 1, 3, 7, 14 days before the end date take the last price at most 24 hours old, keeping only markets still open and 0.02 < p < 0.98. 4) Bin prices and compare each bin's price with the realised Yes frequency, with event-clustered standard errors. 5) Simulate buy No at p <= 0.10 and buy Yes at p >= 0.90 to resolution with 1c or 2c half-spread and the fee in fee-enabled markets; split by end date 60/20/20.

**Grade:** COMPLETE

---
AI-generated research, checked but not peer-reviewed. Historical, experimental research, not investment advice. Dataset Terms of Use: https://alphanotfound.eu/terms